Rapid detection of forest fires is crucial to reduce their devastating impact on ecosystems and human lives. In this paper, we present an AI-based solution for forest fire detection using deep learning from satellite imagery using the ResNet50V2 convolutional neural network (CNN). The dataset used to train the model consists of 1,900 images (950 per class), carefully curated to reflect real-world scenarios of both active forest fires and undisturbed forests. Data preprocessing included image augmentation to reduce overfitting and enhance model performance. Transfer learning, model regularization, and reconstructed pooling layers were applied during training on this dataset, which was augmented with techniques such as random horizontal rotations, zooming, and cropping to improve model generalization. The model achieved 97.63% accuracy and 98.40% precision in detection. Forest fire detection using satellite images is very useful because CNN methods can detect and locate active fires more than once per hour. It is well known that the earlier a forest fire is detected, the more effective it is for people and the environment. This method can help to develop of new strategies for real-time fire monitoring systems, in addition to greatly enhancing wildfire management and prevention efforts. This study focuses not on early fire detection, but on identifying post-wildfire damage using deep learning techniques applied to satellite imagery.
A forest is a closed area of trees, other plant species and animals at a certain level of closure, together
with the invisible is defined as a living system and community in which organisms interact [
Deep learning, as a subset of AI, has the ability to enhance the detection rate of fires and other natural
disasters using large datasets [
ResNet50v2[
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
overall accuracy [
Wildfires can have a devastating effect on ecosystems, people, and economies especially in areas vulnerable to wildfires. Current methods of fire remote sensing via satellite still struggle with offering timeliness, precision, or flexibility. Most of the traditional approaches depend on systematic monitoring (by humans) or use local sensor networks. This aids in scaling detection but isn't helpful in real-time detection over extensive areas. An AI-based method that utilizes satellite images to quickly pinpoint the exact location of the wildfires and assist with rapid aid is highly essential.
For this study, we used the Forest Fire Detection Dataset presented by Khan and Hassan[
The purpose of this paper is to recommend an AI-based approach for fast and effective detection of forest fires. In this regard, the authors trained the ResNet50v2 model which was prepared on a large dataset for forest fire detection and evaluated the model’s performance. The focus of this research is to find possible extensions to current fire detection systems and to emphasize the use of AI in the management of environmental threats.
Over the past few years, there has been significant attention towards the applications of AI and DL on
detecting forest fires [
Harkat et al. and Yang et al. [
Achieving efficient and fast operation of endpoint devices is one of the achievements of Peng et al. [18] with their proposed fire detection algorithm. An effective balance between accuracy and speed is achieved by using quantization-compatible activation functions, a QARep component, and image size optimization using a YOLOv8 algorithm. While image transfer has a positive impact on accuracy, there is an impact on accuracy with respect to INT8 quantization, resulting in some loss of accuracy. The study by Ginkal et al. [19] explores the use of AI methods for forest fire detection. The study provides an AI-based framework for early detection of forest fires. The framework uses machine learning techniques to perform fire detection by combining color, motion and shape features. Features such as color probabilities, color histograms and image moments are used for fire region segmentation, classification, and verification. Experiments show that the proposed framework works with high accuracy and provides real-time processing time.
Another research article by Titu et al. [20] explores the integration of lightweight DL models for real-time fire detection using drones and edge computing. Using knowledge distillation techniques, the study develops DL models such as Detection Transformer (DETR), Detectron2, and YOLOv8. Using this approach, the YOLOv8n model achieved the highest accuracy (95.21%). In another study in a similar area, Anh et al. [21] offer a different approach to detecting forest fires with UAVs, using different color spaces in combination with correlation coefficients to determine the actual fire area.
Liu et al. [22] propose two AI agents armed with large digital databases that autonomously control fiber optic temperature monitoring systems and DL algorithms to detect fires in large commercial spaces. The research examines the effectiveness and reliability of this combined approach and expresses how it can revolutionize fire safety measures when applied to large commercial spaces. In their work, Dampage et al. [23] propose the use of wireless sensor networks in conjunction with machine learning to detect wildfires at very non-extensive stages. Machine learning models are used to evaluate data gathered by sensor networks, so as to estimate the likelihood of a wildfire. In a second part, rechargeable batteries and a solar-powered power supply are used to ensure that the system remains energy efficient.
To summarize, previous research has shown that image-based wildfire detection can be performed with deep learning models VGG, Inception, and variants of YOLO. However, most of these works emphasize detection and monitoring using UAVs or ground sensors. Relatively few have attempted the post-wildfire damage identification using satellite images with high accuracy CNN architectures like ResNet50V2. Our study seeks to fill this gap by utilizing a powerful transfer learning technique for detecting wildfire damage using satellite imagery, providing a valuable resource for post-disaster evaluation and recovery design.
The dataset for forest fire detection is a comprehensive and carefully selected resource specifically designed to assist the development of algorithms for tasks such as forest fire detection and object detection. The images are the result of a search for different keywords in different search engines. As depicted in Fig. 1, designed for the binary problem, the dataset of 1,900 images (950 images per class) is divided into two main categories: The first category contains images documenting active forest fires, while the second category contains images of undisturbed, fire-free forest areas.
To improve the performance of machine learning and DL models, all images in the dataset are three-channel with a spatial resolution of 250 × 250 and consistent formatting. Each image in the dataset is carefully reviewed and pre-processed to remove irrelevant elements, such as human activity or firefighting equipment, to focus only on fire and non-fire regions. This is important for the model that will be used for training, as it eliminates false positives when the model is asked to identify areas of the forest that have burned and those that have not.
This balanced division of the dataset is critical for the model used in training to accurately distinguish between fire-affected and burned areas in forested and unaffected areas. As shown in Table 1, the dataset is divided into three subsets; this separation allows the model to be effectively trained on a variety of samples while achieving higher accuracy rates on the test data.
Augmentation was performed here because the number of data in the dataset is not sufficient for the ResNet architecture and would lead to overfitting of the model. Initially, 20% of the training data was reserved for validation. As shown in Fig. 2, augmentation was then applied at each step of the training: horizontal rotation of the images, random zooming, and certain cropping operations were applied separately for each data in the dataset. For the test data, the data was only scaled to "1./255". Applying the augmentation to the test data may not reflect the real performance of the model and may lead to misleading results.
Convolutional Neural Networks (CNNs) are among the fundamental building blocks of DL methods, and they're popular in tasks such as image and computer vision [24]. CNNs are based on extracting local features from high-order inputs and passing them to lower layers for more complex features [25]. This process allows the model to learn and achieve more accurate results. However, CNNs frequently tend to have issues training deeper, especially in deep networks. This is where deep network architectures like ResNet50v2 can provide a solution. ResNet50v2 is a member of the Residual Networks family and adds an important innovation to the traditional structure of CNNs: residual or jump connections [26]. These structures help solve the problem of gradient loss as the depth of the network increases.
Using ResNet50V2 architecture involves opting for residual blocks which helps to bypass the issues of vanishing and exploding gradient problems during deep representation learning. The purpose of this residual block is captured in an equation that includes the image to be processed, pre-trained weights corresponding to the YOLO CNN, and skip connections. This method is superior at producing results when there are variations in dimension [27]. Furthermore a solution to the degradation is provided using DL framework where the mapping of the layer of the non-linear stack is treated as a ‘cut’ from the original input.
The application of ResNet to computer vision has shown outstanding performance [28]. ResNet18, ResNet50, and ResNet101 are the most widely used types of ResNet networks. Among these network types, ResNet50 has achieved better identification accuracy and real-time performance [29]. The number 50 in its name represents the 50 layers that make up the ResNet50v2 architecture. These layers include the convolutional layer, the batch normalization layer, and the ReLU activation function [30].
In this work, the ResNet50v2 model was transformed into a transfer learning model that was trained on a dataset defined as a forest fire detection dataset. The data was cleaned with fire and non-fire images in an equally weighted ratio. To increase variability and prevent overfitting in the data augmentation stage, random horizontal rotation, zooming, and cropping were applied during the training set hours to allow variability while ensuring that overfitting is contained. The validation and test sets did not require augmentation, although the test set was normalized to a range of 1/255.
The model was then modified to have a global mean pooling layer, a fully connected layer consisting of 128 neurons, an L2 regularization term with λ = 0.01, and a drop of 50%. The binary classification output layer had a sigmoid activation function reflecting the probability of firing. Training was done with the Adam optimizer. The learning rate was set to 0.0001 and the binary cross entropy loss was used. Early stopping was implemented to track accuracy loss, and training was terminated after observing 10 consecutive epochs in which there was no improvement.
Most academic research evaluates model performance not by a single measure, but rather by a combination of measures, including accuracy (1), recall (2), precision (3), and F1-score (4). These metrics allow for fair and comprehensive comparisons across tasks, in addition to providing a quantitative measure of model performance [32].
True Positive (TP )+True Negative (TN )+ False Positive ( FP )+ False Negativ(1e)( FN ) , Recall= Precision=
,
F 1 Score=
(2) (3) (4)
There is a need to analyze the working of the model in a detailed manner, which can be done by analyzing certain different parameters. One of these metrics can be a confusion matrix as shown in Fig. 3. The confusion matrix protocol allows the user to more accurately determine the type of classes that have been solved compared to the others. In this way, it can be determined which fire class was correctly recognized and which was more prone to errors.
This article demonstrates a useful case of DL application for an important ecological problem: detecting wildfires. Applying the ResNet50V2 architecture together with small but cleverly augmented dataset results in 97.63\% accuracy and 98.40\% precision in locating wildfire affected regions from satellite images. With regard to transfer learning, the use of already trained ImageNet weights is one of the most notable advantages, as it allows for better results and faster convergence even when there isn’t sufficient training data available. Data augmentation methods like horizontal flips, zooms, and crops help control overfitting, which is a serious problem when working with small datasets. Regularization techniques also improve the model's generalizability.
Still, the error analysis section would improve by providing more detail outside of the confusion matrix analysis. Determining if the true miss-classifications are false positives (areas without fire but marked as fire) or false negatives (burned areas that should have been marked but are not) is the most important part misclassification analysis. Examining those misclassified images could show the flaws in the model and the biases it holds.
Although many researchers use the ResNet50V2 architecture with transfer learning for image classification, our research is different. We customize the model to detect post-wildfire land damage using satellite images, an area not previously investigated. Unlike most prior work on real-time fire or smoke detection, our focus is on identifying areas of fire damage in forests. Moreover, unlike other more sophisticated solutions such as the dual-agent detection system described in [22], our model is less complicated while achieving the same high accuracy, thus better tailored for environments with constrained resources. To address the problem of limited datasets, we applied specific augmentation strategies, dropout, and L2 regularization. These methods help ensure robustness and generalization. The modular architecture and training pipeline enhance central and edge-based fire monitoring practicality.
Furthermore, examining the practical aspects of the proposed methods would greatly increase the impact of the study. For example, in what ways could this model be used with current wildfire monitoring systems? What are the implications for safeguarding the environment, saving money, and improving response time? Trying to answer these questions would reiterate the importance of the topic while enhancing the discussion providently.
Using the ResNet50V2 architecture, the model achieved a remarkable accuracy of 97.63% when classifying satellite images into wildfire and non-wildfire categories. The model's accuracy, coupled with its capacity to spot forest fires, makes it an invaluable asset for prompt fire detection and prevention. Adopting an approach based on DL and satellite imagery provides the opportunity to enhance the detection of fires’ earliest stages, thus enabling quicker, more efficient actions. In addition, these real-time assessments can aid in firefighting efforts on a personal and communal level. The use of satellite information for instant evaluation can have supportive implications in helping onlookers assess the location of fire activity. This is important in determining the location to dispatch firefighting teams to, hence optimizing resource use and reducing damage.
In addition, this study helps refine the general approach to managing forest fires. As the current version of the model improves, further steps can look into testing other augmentation strategies to increase the model's strength, new fine-tuning adjustments to improve performance, or even other neural network designs that are better intended for certain satellite images or regions geography. Such changes would greatly improve the range of applications of the model so that it can be configured to work in varying environmental and geographical regions, including those that are untapped. This development may enhance the capacity to monitor and prevent wildfires in different ecosystems around the world.
During the preparation of this work, the authors used Grammarly in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [17] Best, N., Ott, J., & Linstead, E. J. (2020). Exploring the efficacy of transfer learning in mining image-based software artifacts. Journal of Big Data, 7, 1–10. [18] Peng, R., Cui, C., & Wu, Y. (2025). Real-time fire detection algorithm on low-power endpoint device. Journal of Real-Time Image Processing, 22(1), 29. [19] Ginkal, P. M., & Kalaiselvi, K. (2024). Forest Fire Detection using AI. Grenze International Journal of Engineering & Technology (GIJET), 10. [20] Titu, M. F. S., Pavel, M. A., Michael, G. K. O., Babar, H., Aman, U., & Khan, R. (2024). Real-Time Fire Detection: Integrating Lightweight Deep Learning Models on Drones with Edge Computing.
Drones, 8(9), 483. [21] Anh, N. D., Van Thanh, P., Lap, D. T., Khai, N. T., Van An, T., Tan, T. D., ... & Dinh, D. N. (2022).
Efficient forest fire detection using rule-based Multi-color Space and correlation coefficient for application in Unmanned Aerial Vehicles. KSII Transactions on Internet and Information Systems (TIIS), 16(2), 381–404. [22] Liu, G., Liu, Z., Qu, G., Ren, L., Wang, L., & Yan, M. (2024). Dual-agent intelligent fire detection method for large commercial spaces based on numerical databases and artificial intelligence.
Process Safety and Environmental Protection, 191, 2485–2499. [23] Dampage, U., Bandaranayake, L., Wanasinghe, R., Kottahachchi, K., & Jayasanka, B. (2022). Forest fire detection system using wireless sensor networks and machine learning. Scientific reports, 12(1), 46. [24] Huang, C., & Yang, Y. (2024, April). Gaussian noise image recognition based on convolutional neural networks. In 2024 5th International Conference on Computer Vision, Image and Deep Learning (CVIDL) (pp. 98–101). IEEE. [25] Shinde, P. P., & Shah, S. (2018, August). A review of machine learning and deep learning applications. In 2018 Fourth international conference on computing communication control and automation (ICCUBEA) (pp. 1–6). IEEE. [26] Rassil, A., Chougrad, H., & Zouaki, H. (2022). Augmented graph neural network with hierarchical global-based residual connections. Neural Networks, 150, 149–166. [27] Riyadi, S., Abidin, F. A., & Audita, N. (2024, May). Comparison of ResNet50V2 and MobileNetV2 Models in Building Architectural Style Classification. In 2024 International Conference on Intelligent Systems and Computer Vision (ISCV) (pp. 1-8). IEEE. [28] Li, B., & Lima, D. (2021). Facial expression recognition via ResNet-50. International Journal of
Cognitive Computing in Engineering, 2, 57–64. [29] Shafiq, M., & Gu, Z. (2022). Deep residual learning for image recognition: A survey. Applied
Sciences, 12(18), 8972. [30] Hindarto, D. (2023). Use ResNet50V2 Deep Learning Model to Classify Five Animal Species.
Jurnal JTIK (Jurnal Teknologi Informasi dan Komunikasi), 7(4), 758–768. [31] Mungoli, N. (2023). Adaptive Ensemble Learning: Boosting Model Performance through Intelligent Feature Fusion in DeepNeural Networks. arXiv preprint arXiv:2304.02653.